Artificial Intelligence and Economic Growth Philippe Aghion College de France and LSE Benjamin F. Jones Northwestern University and NBER Charles I. Jones Stanford GSB and NBER September 1, 2017 Version 0.9 Preliminary and Incomplete Abstract This paper considers potential effects of artificial intelligence (A.I.) on economic growth. We start by modeling A.I. as a process where capital replaces labor at an increasing range of tasks and consider this perspective in light of the evidence to date. We further discuss linkages between A.I. and growth as mediated by firm-level considerations, including organization and market structure. Finally, we engage the concepts of singularities and superintelligence that animate many discussions in the machine intelligence community. The goal throughout is to refine a set of critical questions about A.I. and economic growth and help shape an agenda for the field. We are grateful to Adrien Auclert, Sebastian Di Tella, Pete Klenow, Hannes Mahlmberg, and Chris Tonetti for helpful discussion and comments...
2 P. AGHION, B. JONES, AND C. JONES 1. Introduction This paper considers the implications of artificial intelligence for economic growth. Artificial intelligence (A.I.) can be defined as intelligence exhibited by machines or the capability of a machine to imitate intelligent human behavior. 1 These definitions immediately evoke fundamental economic issues: namely, what happens if A.I. allows an ever-increasing number of tasks previously performed by human labor to become automated i.e., performed by machines? Such A.I. may be deployed in the ordinary production of goods and services with potential effects on growth rates and income shares. But A.I. may also change the production of new ideas themselves. In the near term, A.I. may help solve complex problems and save on computation time. A.I. may also facilitate learning and imitation of technologies across firms, sectors, and activities, thus increasing the scope for knowledge externalities but also for businessstealing. A.I. could increase the scope for introducing new product lines; for example, the recent boost in A.I. following the machine learning revolution has helped precipitate the invention of flying drones and advances toward self-driving cars. Eventually, perhaps A.I. will exceed human creativity in inventing new ideas and new technologies, substituting for even the most skilled researchers. In extreme versions, some observers have argued that A.I. can become rapidly self-improving, producing singularities that feature unbounded machine intelligence and/or unbounded economic growth in finite time (Good (1965), Vinge (1993), Kurzweil (2005)). In this paper we speculate on how A.I. might affect the growth process. Our primary goal here is to help shape an agenda for future research. To do so, we study several growth frameworks and various subtopics, all of which will bear on the implications of A.I. We are concerned with the following questions: How can A.I. affect economic growth when treated as a process of increasing automation in the production of goods and services? Can we reconcile the advent of A.I. with the Kaldor facts, in particular the observed constancy in growth rates and capital share over most of the 20 th century? Should we expect such constancy to persist in the 21 st century given ongoing A.I. advances? 1 The former definition comes for the Wikipedia Artificial Intelligence page and the latter from the Miriam-Webster dictionary.
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 3 How does A.I. affect the internal organization of firms, including skill composition and wage inequality? How does A.I. affect the technology to produce ideas? How may this interface with market structure and firm dynamics? Can A.I. drive massive increases in growth rates, or even a singularity, as some observers predict? Under what conditions, and are these conditions plausible? The paper proceeds as follows. In Section 2, we consider how automation in the production of final goods impacts economic growth. We show that an important indicator of automation is the capital share, so Section 3 examines empirical evidence on capital shares in various sectors of the U.S. and European economies. Section 4 considers how A.I. affects firms, with particular attention to organization, skill composition and wage inequality. Next, Section 5 discusses the possible effects of automation and A.I. on the production of new ideas and knowledge in the context of innovation-led growth, and Section 6 takes this further to consider the possibilities of superintelligence and singularities. Finally, Section 7 concludes by laying out productive directions for further research on A.I. and economic growth. 2. A.I. and Automation of Production One way of looking at the last 150 years of economic progress is that it is driven by automation. The industrial revolution used steam and then electricity to automate many production processes. Relays, transistors, and semiconductors continued this trend. Perhaps artificial intelligence is the next phase of this process rather than a discrete break. It may be a natural progression from autopilots, computer-controlled automobile engines, and MRI machines to self-driving cars and A.I. radiology reports. An advantage of this perspective is that it allows us to use historical experience to inform us about the possible future effects of A.I. Later sections will explore alternatives that do not make this assumption.
4 P. AGHION, B. JONES, AND C. JONES 2.1 The Zeira (1998) Model of Automation and Growth A clear and elegant model of automation is the task-based model of Zeira (1998). In its simplest form, Zeira considers a production function like Y = AX α 1 1 Xα 2 2... Xn αn where n α i = 1. (1) i=1 Tasks that have not yet been automated can be produced one-for-one by labor. Once a task is automated, one unit of capital can be used instead: X i = L i K i if not automated (2) if automated If the aggregate capital K and labor L are assigned to these tasks optimally, the production function can be expressed (up to an unimportant constant) as Y = AK α L 1 α (3) where it is now understood that the exponent α reflects the overall share and importance of tasks that have been automated. Next, we embed this setup into a standard neoclassical growth model with a constant investment rate; in fact, for the remainder of the paper this is how we will close the capital/investment side of the model for simplicity. The share of factor payments going to capital is given by α and the long-run growth rate of y Y/L is g y = g 1 α, (4) where g is the growth rate of A. An increase in automation will therefore increase the capital share α and, because of the multiplier effect associated with capital accumulation, increase the long-run growth rate. Zeira emphasizes that automation has been going on at least since the industrial revolution, and his elegant model helps us to understand that. However, its strong predictions that growth rates and capital shares should be rising with automation go against the famous Kaldor (1961) stylized facts that growth rates and capital shares are
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 5 relatively stable over time. In particular, this stability is a good characterization of the U.S. economy for the bulk of the 20th century; for example, see Jones (2016). The Zeira framework, then, needs to be improved so that it is consistent with historical evidence. Acemoglu and Restrepo (2016) provide one approach to solving this problem. They allow CES production and endogenize the number of tasks. In particular, they suppose that research can take two different directions: discovering how to automate an existing task or discovering new tasks that can be used in production. In their setting, α reflects the fraction of tasks that have been automated. This leads them to emphasize one possible resolution to the empirical shortcoming of Zeira: perhaps we are inventing new tasks just as quickly as we are automating old tasks. So the fraction of tasks that are automated is constant, leading to a stable capital share and a stable growth rate. Other important contributions to this rapidly expanding literature include Peretto and Seater (2013) and Hemous and Olsen (2016). Peretto and Seater (2013) explicitly consider a research technology that allows firms to change the exponent in a Cobb- Douglas production function; while they do not emphasize the link to the Zeira model, with hindsight the connections to that approach to automation are interesting. The model of Hemous and Olsen (2016) is closely related to what follows in the next subsection. They focus on CES production instead of Cobb-Douglas, as we do below, but emphasize the implications of their framework for wage inequality between highskilled and low-skilled workers. The next section takes a complementary approach, building on this literature and using the insights of Zeira and automation to understand the structural change associated with Baumol s cost disease. 2.2 Automation and Baumol s Cost Disease 2.2.1 Overview The share of agriculture in GDP or employment is falling toward zero. The same is true for manufacturing in many countries of the world. Maybe automation increases the capital share in these sectors and also interacts with nonhomotheticities in production or consumption to drive the GDP shares toward zero. The aggregate capital share is then a balance of a rising capital share in agriculture/manufacturing/automated goods with a declining GDP share of these goods in the economy.
6 P. AGHION, B. JONES, AND C. JONES Looking toward the future, 3D-printing techniques and nanotechnology that allow production to start at the molecular or even atomic level could someday automate all manufacturing. Could A.I. do the same thing in many service sectors? What would economic growth look like in such a world? This section expands on the Zeira (1998) and Acemoglu and Restrepo (2016) models to develop a framework that is consistent with the large structural changes in the economy. Baumol (1967) observed that rapid productivity growth in some sectors relative to others could result in a cost disease in which the slow growing sectors become increasingly important in the economy. We explore the possibility that automation is the force behind these changes. 2.2.2 Model GDP is a CES combination of goods with an elasticity of substitution less than one: Y t = A t ( 1 0 Y ρ it di ) 1/ρ where σ 1 1 ρ < 1 (5) where A t = A 0 e gt captures standard technological change, which we take to be exogenous for now. As in Zeira, another part of technical change is the automation of production. Goods that have not yet been automated can be produced one-for-one by labor. When a good has been automated, one unit of capital can be used instead: Y it = L it K it if not automated (6) if automated This division is stark to keep the model simple. An alternative would be to say that goods are produced with a Cobb-Douglas combination capital and labor, and when a good is automated, it is produced with a higher exponent on capital. 2 The remainder of the model is neoclassical: Y t = C t + I t (7) 2 A technical condition is required, of course, so that tasks that have been automated are actually produced with capital instead of labor. We assume this condition holds.
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 7 K t = I t δk t (8) 1 0 1 We assume a fixed endowment of labor for simplicity. 0 K it di = K t (9) L it di = L (10) Let β t be the fraction of goods that that have been automated as of date t. Then, with symmetry, the production function can be written as Y t = A t ( β 1 ρ t K ρ t + (1 β t) 1 ρ L ρ) 1/ρ. (11) This setup therefore reduces to a particular version of the neoclassical growth model, and the allocation of resources can be decentralized in a standard competitive equilibrium. In this equilibrium, the share of automated goods in GDP equals the share of capital in factor payments: α Kt Y t K t K t Y t = β 1 ρ t A ρ t ( Kt Y t ) ρ. (12) Similarly, the share of non-automated goods in GDP equals the labor share of factor payments: α Lt Y t L t = β 1 ρ t L t Y t A ρ t ( Lt Y t ) ρ. (13) And therefore the ratio of automated to nonautomated output or the ratio of the capital share to the labor share equals ( ) α 1 ρ ( ) ρ Kt βt Kt =. (14) α Lt 1 β t L t Finally, notice that the production function in equation (11) is just a special case of a neoclassical production function: 1 ρ ρ Y t = A t F (B t K t, C t L t ) where B t βt and C t (1 β) 1 ρ ρ. (15) With ρ < 0, notice that β t B t and C t. That is, automation is equivalent to a combination of labor-augmenting technical change and capital-depleting technical
8 P. AGHION, B. JONES, AND C. JONES change. This is surprising. One might have thought of automation as somehow capital augmenting. Instead, it is very different: it is labor augmenting and simultaneously dilutes the stock of capital. Notice that these conclusions would be reversed if the elasticity of substitution were greater than one; they importantly rely on ρ < 0. This opens up one possibility that we will explore below: what happens if the evolution of β t is such that C t grows at a constant exponential rate? This can occur if 1 β t falls at a constant exponential rate toward zero, meaning that β t 1 in the limit and the economy gets ever closer to full automation (but never quite reaches that point). The logic of the neoclassical growth model suggests that this could produce a balanced growth path with constant factor shares, at least in the limit. 3 2.2.3 Discussion The last several equations have a number of implications that can now be explored. First, we specified from the beginning that we are interested in the case in which the elasticity of substitution between goods is less than one, so that ρ < 0. From equation (14), there are two basic forces that move the capital share (or, equivalently, the share of the economy that is automated). First, an increase in the fraction of goods that are automated, β t, will increase the share of automated goods in GDP and increase the capital share (holding K/L constant). This is intuitive and repeats the logic of the Zeira model. Second, as K/L rises, the capital share and the value of the automated sector as a share of GDP will decline. Essentially, with an elasticity of substitution less than one, the price effects dominate. The price of automated goods declines relative to the price of non-automated goods because of capital accumulation. Because demand is relatively inelastic, the expenditure share of these goods declines as well. Automation and Baumol s cost disease are then intimately linked. Perhaps the automation of agriculture and manufacturing leads these sectors to grow rapidly and causes their shares in GDP to decline. 4 The bottom line is that there is a race between these two forces. As more sectors are automated, β increases, and this tends to increase the share of automated goods and capital. But because these automated goods experience faster growth, their price 3 This requires A t to be constant. 4 Manuelli and Seshadri (2014) offer a systematic account of the how the tractor gradually replaced the horse and in American agriculture between 1910 and 1960.
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 9 declines, and the low elasticity of substitution means that their shares of GDP also decline. One could imagine, following Acemoglu and Restrepo (2016), writing down a technology by which research effort leads goods to be automated. But it is relatively clear that depending on exactly how one specifies this technology, β t 1 β t can rise faster or slower than (K t /L t ) ρ declines. That is, the result would depend on detailed assumptions related to automation, and we do not have strong priors on how to make these assumptions. We leave this to future work and focus for now on what happens when β t changes in different ways. 2.2.4 Balanced Growth (Asymptotically) Recall that the production function in this economy can be written in factor-augmenting form as 1 ρ ρ Y t = F (B t K t, C t L t ) where B t βt and C t (1 β) 1 ρ ρ. (16) In this section, we explicitly omit any form of technical change other than automation and show how automation can produce a balanced growth path asymptotically. particular, we want to consider an exogenous time path for the fraction of tasks that are automated, β t, such that β t 1 but in a way that C t grows at a constant exponential rate. This turns out to be straightfoward. Let γ t 1 β t, so that C t = γ 1 ρ ρ t In. Because the exponent is negative (ρ < 0), if γ falls at a constant exponential rate, C t will grow at a constant exponential rate. This occurs if β t = θ(1 β t ), implying that g γ = θ. Intuitively, a constant fraction, θ, of the tasks that have not yet been automated become automated each year. Figure 1 shows that this example can produce steady exponential growth. We begin in year 0 with none of the goods being automated, and then have a constant fraction of the remainder being automated each year. There is obviously enormous structural change underlying and generating the stable exponential growth of GDP in this case. The capital share of factor payments begins at zero and then rises gradually over time, eventually asymptoting to a value around 1/3. Even though an ever-vanishing fraction of the economy has not yet been automated, so labor has less and less to do, the fact that automated goods are produced with cheap capital combined with an elasticity of substitution less than one means that the automated share of GDP remains at 1/3
10 P. AGHION, B. JONES, AND C. JONES and labor still earns around 2/3 of GDP asymptotically! Along such a path, however, sectors like agriculture and manufacturing exhibit a structural transformation. For example, let sectors on the interval [0, 1/3] denote agriculture and the automated portion of manufacturing as of some year, such as 1990. These sectors experience a declining share of GDP over time, as their prices fall rapidly. The automated share of the economy will be constant only because new goods are becoming automated. 2.2.5 Constant Factor Shares Another interesting case worth considering is under what conditions can this model produce factor shares that are constant over time? Taking logs and derivatives of (14), the capital share will be constant if and only if ( ) ρ g βt = (1 β t ) g kt, (17) 1 ρ where g kt is the growth rate of k K/L. This is very much a knife-edge condition. It requires the growth rate of β t to slow over time at just the right rate as more and more goods get automated. Figure 2 shows an example with this feature, in an otherwise neoclassical model with exogenous growth in A t at 2% per year. That is, unlike the previous section, we allow other forms of technological change to make tractors and computers better of time, in addition to allowing automation. In this simulation, automation proceeds at just the right rate so as to keep the capital share constant for the first 150 years. After that time, we simply assume that β t is constant and automation stops, so as to show what happens in that case as well. The perhaps surprising result in this example is that the constant factor shares occur while the growth rate of GDP rises at an increasing rate. From the earlier simulation in Figure 1, one might have inferred that a constant capital share would be associated with declining growth. However, this is not the case and instead growth rates increase. The key to the explanation is to note that with some algebra, we can show that the constant factor share case requires g Y t = g A + β t g Kt. (18)
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 11 3% Figure 1: Automation and Asymptotic Balanced Growth GROWTH RATE OF GDP 2% 1% 0% 0 50 100 150 200 250 300 350 400 450 500 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 (a) The Growth Rate of GDP over Time Fraction automated, t Capital share YEAR 0 0 50 100 150 200 250 300 350 400 450 500 (b) Automation and the Capital Share K YEAR Note: This simulation assumes ρ < 0 and that a constant fraction of the tasks that have not yet been automated become automated each year. Therefore C t (1 β) 1 ρ ρ grows at a constant exponential rate (2% per year in this example), leading to an asymptotic balanced growth path. The share of tasks that are automated approaches 100% in the limit. Interestingly, the capital share of factor payments (and the share of automated goods in GDP) remains bounded, in this case at a ( value ) around 1/3. With a constant investment rate of s, the limiting value of the capital share is s ρ. g Y +δ
12 P. AGHION, B. JONES, AND C. JONES GROWTH RATE OF GDP 5% Figure 2: Automation with a Constant Capital Share 4% 3% 2% 0 50 100 150 200 250 300 (a) The Growth Rate of GDP over Time YEAR 0.7 0.6 Fraction automated, t 0.5 0.4 0.3 0.2 0.1 Capital share K 0 0 50 100 150 200 250 300 (b) Automation and the Capital Share YEAR Note: This simulation assumes ρ < 0 and sets β t so that the capital share is constant between year 0 and year 150. After year 150, we assume β t stays at its constant value. A t is assumed to grow at a constant rate of 2% per year throughout.
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 13 First, consider the case with g A = 0. We know that a true balanced growth path requires g Y = g K. This can occur in only two ways if g A = 0: either β t = 1 or g Y = g K = 0 if β t < 1. The first case is the one that we explored in the previous example back in Figure 1. The second case shows that if g A = 0, then constant factor shares will be associated with zero exponential growth. Now we can see the reconciliation between Figures 1 and 2. In the absence of g A > 0, the growth rate of the economy would fall to zero. Introducing g A > 0 with constant factor shares does increases the growth rate. To see why growth has to accelerate, equation (18) is again useful. If growth were balanced, then g Y = g K. But then the rise in β t would tend to raise g Y and g K. This is why growth accelerates. 2.2.6 Regime Switching A final simulation shown in Figure 3 combines aspects of the two previous simulations to produce results closer in spirit to our observed data, albeit in a highly stylized way. We assume that automation alternates between two regimes. The first is like Figure 1, in which a constant fraction of the remaining tasks are automated each year, tending to raise the capital share and produce high growth. In the second, β t is constant and no new automation occurs. In both regimes, A t grows at a constant rate of 0.4% per year, so that even when the fraction of tasks being automated is stagnant, the nature of automation is improving, which tends to depress the capital share. Regimes last for 30 years. Period 100 is highlighted with a black circle. At this point in time, the capital share is relatively high and growth is relatively low. By playing with parameter values, including the growth rate of A t and β t, it is possible to get a wide range of outcomes. For example, the fact that the capital share in the future is lower than in period 100 instead of higher can be reversed. 2.2.7 Summing Up Automation an increase in β t can be viewed as a twist of the capital- and laboraugmenting terms in a neoclassical production function. From Uzawa s famous theorem, since we do not in general have purely labor-augmenting technical change, this setting will not lead to balanced growth. In this particular application (e.g. with ρ < 0), either the capital share or the growth rate of GDP will tend to increase over time,
14 P. AGHION, B. JONES, AND C. JONES GROWTH RATE OF GDP 3% Figure 3: Intermittent Automation to Match Data? 2% 1% 0% 0 50 100 150 200 250 300 1 0.9 (a) The Growth Rate of GDP over Time YEAR 0.8 Fraction automated, t 0.7 0.6 0.5 0.4 0.3 0.2 Capital share K 0.1 0 50 100 150 200 250 300 (b) Automation and the Capital Share YEAR Note: This simulation combines aspects of the two previous simulations to produce results closer in spirit to our observed data. We assume that automation alternates between two regimes. In the first, a constant fraction of the remaining tasks are automated each year. In the second, β t is constant and no new automation occurs. In both regimes, A t grows at a constant rate of 0.4% per year. Regimes last for 30 years. Period 100 is highlighted with a black circle. At this point in time, the capital share is relatively high and growth is relatively low.
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 15 and sometimes both. We showed one special case in which all tasks are ultimately automated that produced balanced growth in the limit with a constant capital share less than 100%. A shortcoming of this case is that it requires automation to be the only form of technological change. If, instead, the nature of automation itself improves over time consider the plow, then the tractor, then the combine-harvester, then GPS tracking then the model is best thought of as featuring both automation and something like improvements in A t. In this case, one would generally expect growth not to be balanced. However, a combination of periods of automation followed by periods of respite, like that shown in Figure 3 does seem capable of producing dynamics at least superficially similar to what we ve seen in the U.S. in recent years: a period of a high capital share with relatively slow economic growth. 3. Evidence on Capital Shares and Automation The models of the previous section suggest that a key place to look for evidence on automation is the share of factor payments going to capital the capital share. In recent years, the rise in the capital share in the U.S. and around the world has been a central topic of research. For example, see Karabarbounis and Neiman (2013), Elsby, Hobijn and Şahin (2013), and Kehrig and Vincent (2017). In this section, we explore this evidence, first for industries within the United States, second for the motor vehicles industry in the U.S. and Europe, and finally by looking at how changes in capital shares over time correlate with the adoption of robots. Figure 4 reports capital shares by industry from the U.S. KLEMS data of Jorgenson, Ho and Samuels (2017); shares are smoothed using an HP filter with smoothing parameter 400 to focus on the medium- to long-run trends. It is well-known that the aggregate capital share has increased since at least the year 2000 in the U.S. economy. Figure 4 shows that this aggregate trend holds up across a large number of sectors, including agriculture, construction, chemicals, computers equipment manufacturing, motor vehicles, publishing, telecommunications, and wholesale and retail trade. The main place where one does not see this trend is in services, including education, government, and health. In those sectors, the capital share is relatively stable or perhaps increasing slightly since 1990. But the big trend one sees in these data from services is
16 P. AGHION, B. JONES, AND C. JONES Figure 4: U.S. Capital Shares by Industry 1 Petroleum Mfg. 1 0.8 Oil/Gas Extraction Utilities 0.8 Chemicals 0.6 Agriculture 0.6 Motor Vehicles 0.4 0.2 Construction 0.4 0.2 Computers Plastics Furniture 0 1940 1960 1980 2000 2020 1 0.8 0.6 0.4 0.2 Air Trans. Movies Publishing Wholesale Retail 0 1940 1960 1980 2000 2020 0 1940 1960 1980 2000 2020 1 0.8 0.6 0.4 0.2 Health (hospitals) Telecommunications Federal Govt Health (ambulatory) Education 0 1940 1960 1980 2000 2020 Note: The graph reports capital shares by industry from the U.S. KLEMS data of Jorgenson, Ho and Samuels (2017). Shares are smoothed using an HP filter with smoothing parameter 400. a large downward trend between 1950 and 1980. It would be interesting to know more about what accounts for this trend. While the facts are broadly consistent with automation (or an increase in automation), it is also clear that capital and labor shares involve many other economic forces as well. For example, Autor, Dorn, Katz, Patterson and Van Reenen (2017) suggest that a composition effect involving a shift toward superstar firms with high capital shares underlies the industry trends. That paper and Barkai (2017) propose that a rise in industry concentration and markups may underlie some of the increases in the capital share. Changes in unionization over time may be another contributing factor to the dynamics of factor shares. This is all to say that a much more careful analysis of factor shares and automation is required before any conclusions can be drawn. Keeping that important caveat in mind, Figure 5 shows evidence on the capital share in the motor vehicles manufacturing industry for the U.S. and several European countries. As Acemoglu and Restrepo (2017) note (more on this below), the motor
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 17 CAPITAL SHARE 55 Figure 5: The Capital Share for Motor Vehicles 50 45 40 35 30 25 Germany U.S. Spain Italy 20 15 France U.K. 10 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 Note: Data for the European countries are from the EU-KLEMS project at http://www.euklems.net/ for the transportation equipment sector; see Jägger (2016). U.S. data are from Jorgenson, Ho and Samuels (2017). Shares are smoothed using an HP filter with smoothing parameter 400. YEAR vehicles industry is by far the industry that has invested most heavily in industrial robots during the past two decades, so this industry is particularly interesting from the standpoint of automation. The capital share in motor vehicles shows a large increase in the United States, France, Germany, and Spain in recent decades. Interestingly, Italy and the U.K. exhibit declines in the capital share for motor vehicles since 1995. The absolute level differences in the capital share for motor vehicles in 2014 are also interesting, ranging from a high of more than 50 percent in the U.S. to a low of around 20 percent in recent years in the U.K. Clearly it would be valuable to better understand these large differences in levels and trends. Automation is likely only a part of the story. Acemoglu and Restrepo (2017) use data from the International Federation of Robots to study the impact of the adoption of industrial robots on the U.S. labor market. At the industry level, this data is available for the decade 2004 to 2014. Figure 6 shows data on the change in capital share by industry versus the change in the use of industrial robots. Two main facts stand out from the figure. First, as noted earlier, the motor vehicles
18 P. AGHION, B. JONES, AND C. JONES CHANGE IN CAPITAL SHARE 0.2 Figure 6: Capital Shares and Robots, 2004 2014 Other Transport Equipment 0.15 0.1 0.05 0 Machinery Appliances Chemicals Misc Mfg Computers Paper Food Mfg Textiles Wood Plastics Utilities Minerals Primary Metals Fabricated Metals Construction Education Motor Vehicles -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 CHANGE IN ROBOTS/VA Note: The graph plots the change in the capital share from Jorgenson, Ho and Samuels (2017) against the change in the stock of robots relative to value-added using the robots data from Acemoglu and Restrepo (2017). industry is by far the largest adopter of industrial robots. For example, more than 56 percent of new industrial robots purchased in 2014 were installed in the motor vehicles industry; the next highest share was under 12 percent in computers and electronic products. Second, there is little correlation between automation as measured by robots and the change in the capital share between 2004 and 2014. The overall level of industrial robot penetration is relatively small, and as we discussed earlier, other forces including changes in market power, unionization, and composition effects are moving capital shares around in a way that makes it hard for a simple data plot to disentangle. 4. A.I. and firms: organization, skills and wage inequality How should we expect firms to adapt their internal organization, the skill composition of their workforce and their wage policies to the introduction of AI? In his recent book on The Economics of the Common Good, Tirole (2017) spells out what one may consider to be common wisdom expectations on firms and A.I. Namely, introducing A.I.
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 19 should: (a) increase the wage gap between skilled and unskilled labor, as the latter is presumably more substitutable to A.I. than the former; (b) the introduction of A.I. allows firms to automate and dispense with middle-men performing monitoring tasks (in order words, firms should become flatter, i.e. with higher spans of control); (c) should encourage self-employment by making it easier for individuals to build up reputation. Let us revisit these various points in more details. A.I., skills, and wage premia: On A.I. and the increased gap between skilled and unskilled wage, the prediction brings us back to Krusell, Ohanian, Ríos-Rull and Violante (2000): based on an aggregate production function in which physical equipment is more substitutable to unskilled labor than to skilled labor, these authors argued that the observed acceleration in the decline of the relative price of production equipment goods since the mid-1970s could account for most of the variation in the college premium over the past twenty-five years. In other words, the rise in the college premium could largely be attributed to an increase in the rate of (capital-embodied) skill-biased technical progress. And presumably A.I. is an extreme form of capital-embodied skill-biased technical change, as robots substitute for unskilled labor but require skilled labor to be installed and exploited. However, recent work by Aghion, Bergeaud, Blundell and Griffith (2017) suggests that while the prediction of a premium to skills may hold at the macroeconomic level, it perhaps misses important aspects of firms internal organization and that organization itself may evolve as a result of introducing A.I. More specifically, Aghion, Bergeaud, Blundell and Griffith (2017) use matched employer-employee data from the UK, which they augment with information on R&D expenditures, to analyze the relationship between innovativeness and average wage income across firms. A first, not surprising, finding is that more R&D intensive firms pay higher wages on average and employ a higher fraction of high-occupation workers than less R&D intensive firms (see Figure 7 below). This, in turn, is perfectly in line with the above prediction (a) but also with prediction (b) as it suggests that more innovative (or more frontier ) firms rely more on outsourcing for low-occupation tasks. However, a more surprising finding in Aghion, Bergeaud, Blundell and Griffith (2017) is that lower-skilled (lower occupation) workers benefit more from working in more R&D intensive firms (relative to working in a firm which does no R&D) than higher-skilled workers. This finding is summarized by Figure
20 P. AGHION, B. JONES, AND C. JONES Figure 7: Log hourly wage and R&D intensity Note: This figure plots the logarithm of total hourly income against the logarithm of total R&D expenditures (intramural + extramural) per employee (R&D intensity). Source: Aghion, Bergeaud, Blundell and Griffith (2017).
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 21 Figure 8: Log hourly wage and R&D intensity Note: This figure plots the logarithm of total hourly income against the logarithm of total R&D expenditures (intramural + extramural) per employee (R&D intensity) for different skill groups. Source: Aghion, Bergeaud, Blundell and Griffith (2017). 8. In that Figure, we first see that higher-skilled workers earn more than lower-skilled workers in any firm no matter how R&D intensive that firm is (the high-skill wage curve always lies strictly above the middle-skill curve which itself always lies above the lowerskill curve). But more interestingly the lower-skill curve is steeper than the middleskill and higher-skill curve. But the slope of each of these curves precisely reflects the premium for workers with the corresponding skill level to working in a more innovative firm. Similarly, we should expect more AI-intensive firms to: (i) employ a higher fraction of (more highly paid) high-skill workers; (ii) outsource an increasing fraction of lowoccupation tasks; (iii) give a higher premium to those low-occupation workers they keep within the firm (unless we take the extreme view that all the functions to be per-
22 P. AGHION, B. JONES, AND C. JONES formed by low-occupation workers could be performed by robots). To rationalize the above findings and these latter predictions, let us follow Aghion, Bergeaud, Blundell and Griffith (2017) who propose a model in which more innovative firms display a higher degree of complementarity between low-skill workers and the other production factors (capital and high-skill labor) within the firm. Another feature of their model is that high-occupation employees skills are less firm-specific than lowskill workers: namely, if the firm was to replace a high-skill worker by another high-skill worker, the downside risk would be limited by the fact that higher-skill employees are typically more educated employees, whose market value is largely determined by their education and accumulated reputation, whereas low-occupation employees quality is more firm-specific. This model is meant to capture the idea that low-occupation workers can have a potentially more damaging effect on the firm s value if the firm is more innovative (or more A.I. intensive for our purpose). In particular an important difference with the common wisdom, is that here innovativeness (or A.I. intensity) impacts on the organizational form of the firm and in particular on complementarity or substitutability between workers with different skill levels within the firm, whereas the common wisdom view takes this complementarity or substitutability as given. Think of a low-occupation employee (for example an assistant) who shows outstanding ability, initiative and trustworthiness. That employee performs a set of tasks for which it might be difficult or too costly to hire a high-skill worker; furthermore, and perhaps more importantly, the low-occupation employee is expected to stay longer in the firm than higher-skill employees, which in turn encourages the firm to invest more in trust-building and firm-specific human capital and knowledge. Overall, such low-occupation employees can make a big difference to the firm s performance. This alternative view of A.I. and firms, is consistent with the work of theorists of the firm such as Luis Garicano. Thus in Garicano (2000) downstream - low-occupation - employees are consistently facing new problems; among these new problems they sort out those they can solve themselves (the easier problems) and the more difficult questions they pass on to upstream higher-skill - employees in the firm s hierarchy. Presumably, the more innovative or more A.I. intensive- the firm is, the harder it is to solve the more difficult questions, and therefore the more valuable the time of up-
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 23 stream high-occupation employees becomes; this in turn makes it all the more important to employ downstream - low-occupation - employees with higher ability to make sure that less problems will be passed on to the upstream - high-occupation - employees within the firm, so that these high-occupation employees will have more free time to concentrate on solving the most difficult tasks. Another interpretation of the higher complementarity between low-occupation and high-occupation employees in more innovative (or more AI-intensive) firms, is that the potential loss from unreliable low-occupation employees is bigger in such firms: hence the need to select out those low-occupation employees which are not reliable. This higher complementarity between low-occupation workers and other production factors in more innovative (or more A.I. intensive) firms in turn increases the bargaining power of low-occupation workers within the firm (it increases their Shapley Value if we follow Stole and Zwiebel (1996)). This in turn explains the higher payoff for low-occupation workers. It also predicts that job turnover should be lower (tenure should be higher) amongst lowoccupation workers who work for more innovative (more AI-intensive) firms than for low-occupation workers who work for less innovative firms, whereas the turnover difference should be less between high-occupation workers employed by these two types of firms. This additional prediction is also confronted to the data in Aghion, Bergeaud, Blundell and Griffith (2017). Note that so far R&D investment has been used as the measure of the firm s innovativeness or frontierness. We would like to test the same predictions but using explicit measures of A.I. intensity as the RHS variable in the regressions (investment in robots, reliance on digital platforms,..). A.I. and firm s organizational form: Recent empirical studies (e.g. see Bloom, Garicano, Sadun and Van Reenen (2014)), have shown that the IT revolution has led firms to eliminate middle-range jobs and move towards flatter organizational structure. The development of A.I. should reinforce that trend, while perhaps also reducing the ratio to low-occupation to high-occupation jobs within firms as we argued above. A potentially helpful framework to think about firms organizational forms, is Aghion and Tirole (1997). There, a principal can decide whether or not to delegate authority to a downstream agent. She can delegate authority in two ways: (i) by formally allocating control rights to the agent (in that case we say that the principal delegates formal
24 P. AGHION, B. JONES, AND C. JONES authority to the agent); (ii) or informally through the design of the organization, e.g. by increasing the span of control or by engaging in multiple activities: these devices enable the principal to commit to leave initiative to the agent (in that case we say that the principal delegates real authority to the agent). And agents initiative particularly matters if the firm needs to be innovative, which is particularly the case for more frontier firms in their sectors. Whether she decides to delegate formal or only real authority to her agent, the principal faces the following trade-off: more delegation of authority to the agent induces the agent to take more initiative; on the other hand this implies that the principal will lose some control over the firm, and therefore face the possibility that suboptimal decisions (from her viewpoint) be taken more often. Which of these two counteracting effects of delegation dominates, will in turn depend upon the degree of congruence between the principal s and the agent s preference, but also about the principal s ability to reverse suboptimal decisions. How should the introduction of A.I. affect this trade-off between loss of control and initiative? To the extent that A.I. makes it easier for the principal to monitor the agent, more delegation of authority will be required in order to still elicit initiative from the agent. The incentive to delegate more authority to downstream agents, will also be enhanced by the fact that with A.I., suboptimal decision-making by downstream agents can be more easily corrected and reversed: in other words, A.I. should reduce the loss of control involved in delegating authority downstream. A third reason for why A.I. may encourage decentralization in decision-making, has to do with coordination costs: namely, it may be costly for the principal to delegate decision making to downstream units if this prevents these units from coordinating within the firm (see Hart and Holmstrom (2010)). But here again, A.I. may help overcome this problem by reducing the monitoring costs between the principal and its multiple downstream units, and thereby induce more decentralization of authority. More delegation of authority in turn can be achieved through various means: in particular by eliminating intermediate layers in the firm s hierarchy, or by turning downstream units into profit centers or fully independent firms, or through horizontal integration which will commit the principal to spending time on other activities. Overall, one can imagine that the development of A.I. in more frontier sectors should lead to larger and more horizontally integrated firms, to flatter firms with more profit centers,
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 25 which outsource an increasing number of tasks to independent self-employed agents. The increased reliance on self-employed independent agents will in turn be facilitated by the fact that, as well explained by Tirole (2017), AI helps agents to quickly develop individual reputations. This brings us to the third aspect of A.I. and organizations on selfemployment. A.I. and self-employment: As stressed above, A.I. favors the development of self-employment for at least two reasons: first, it may induce A.I. intensive firms to outsource tasks, starting with low-occupation tasks; second, it makes it easier for independent agents to develop individual reputations. Does that imply that A.I. should result in the end of large integrated firms with individuals only interacting with each other through platforms? And which agents are more likely to become self-employed? On the first question: Tirole (2017) provides at least two reasons for why firms should survive the introduction of A.I. First, some activities involve large sunk costs and/or large fixed costs that cannot be borne by a single individual. Second, some activities involve a level of risk-taking which also may not be borne by one single agent. To this we should add the transaction cost argument that vertical integration facilitates relationspecific investments in situations of contractual incompleteness: can we truly imagine that A.I. will by itself fully overcome contractual incompleteness? On the second question: Our above discussion suggests that low-skill activities involving limited risk and for which A.I. helps develop individual reputations (hotel or transport services, health assistance to the elder and/or handicapped, catering services, house cleaning,..) are primary candidates for increasingly becoming self-employment jobs as A.I. diffuses in the economy. And indeed recent studies by Saez (2010), Chetty, Friedman, Olsen and Pistaferri (2011), and Kleven and Waseem (2013) point to lowincome individuals being more responsive to tax or regulatory changes aimed at facilitating self-employment. Natural extensions of these studies would be to explore the extent to which such regulatory changes have had more impact in sectors with higher A.I. penetration. The interplay between A.I. and self-employment also involves potentially interesting dynamic aspects. Thus it might be worth looking at whether self-employment helps individuals accumulate human capital (or at least protects them against the risk of human capital depreciation following the loss of a formal job), and the more so in sectors with higher AI penetration. Also interesting would be to look at the interplay
26 P. AGHION, B. JONES, AND C. JONES between self-employment and A.I. is itself affected by government policies and institutions, and here we have primarily in mind education policy and social or income insurance for the self-employed. How do these policies affect the future performance of currently self-employed individuals, and are they at all complemented by the introduction of AI? In particular, do currently self-employed individuals move back to working for larger firms, and how does the probability of moving back to a regular employment vary with A.I., government policy, and the interplay between the two? Presumably, a more performing basic education system and a more comprehensive social insurance system should both encourage self-employed individuals to better take advantage of A.I. opportunities and support to accumulate skills and reputation and thereby improve their future career prospects. On the other hand, some may argue that A.I. will have a discouraging effect on self-employed individuals, if it lowers their prospects of ever reintegrating a regular firm in the future, as more A.I. intensive firms reduce their demand for low-occupation workers. 5. A.I. and Innovation-Based Growth In the previous sections, we examined the implications of introducing A.I. in the production function for goods and services. But what if the tasks of the innovation process themselves can be automated? How would A.I. interact with the production of new ideas? In this section, we first introduce A.I. in the production technology for new ideas and look at how A.I. affects growth. We then consider a other channels through which A.I. could influence growth, including product market competition, cross-sector incentives for innovation, and business-stealing. In general, this section lays the groundwork for the next section, where we consider the possibility that A.I. could lead to a singularity. 5.1 A.I. in the Idea Production Function A moment of introspection into our own research process reveals many ways in which automation can matter for the production of ideas. Research tasks that have benefited from automation and technological change include typing and distributing our papers, obtaining research materials and data (e.g. from libraries), ordering supplies,
ARTIFICIAL INTELLIGENCE AND ECONOMIC GROWTH 27 analyzing data, solving math problems, and computing equilibrium outcomes. Beyond economics, other examples include carrying out experiments, sequencing genomes, exploring various chemical reactions and materials. In other words, applying the same task-based model to the idea production function and considering the automation of research tasks seems relevant. To keep things simple, suppose the production function for goods and services just uses labor and ideas: Y t = A t L t. (19) But suppose that various tasks are used to make new ideas according to ( 1 ) 1/ρ A t = A φ t X ρ it di where σ 1 0 1 ρ < 1 (20) Assuming some fraction β t of tasks have been automated using a similar setup to that in Section 2 the idea production function can be expressed as A t = A φ t ((B tk t ) ρ + (C t S t ) ρ ) 1/ρ A φ t F (B tk t, C t S t ) (21) where S t is the research labor used to make ideas, and B t and C t are defined as before, namely B t β 1 ρ ρ t and C t (1 β t ) 1 ρ ρ. Several observations then follow from this setup. First, consider the case in which β is constant at some value but then increases to a higher value (recall that this leads to a one-time decrease in B and increase in C). The idea production function can then be written as A t = A φ t S tf A φ t CS t ( BKt S t ), C where the notation means is asymptotically proportional to. The second line follows if K t /S t is growing over time (i.e. if there is economic growth) and if the elasticity of substitution in F ( ) is less than one, which we ve assumed. In that case, the CES function is bounded by its scarcest argument, in this case researchers. Automation then essentially produces a level effect but leaves the long-run growth rate of the economy unchanged if φ < 1. Alternatively, if φ = 1 the classic endogenous growth case then automation raises long-run growth. (22)